Neural Persistence

A complexity measure introduced by Rieck, Togninalli, Bock, Moor, Horn, Gumbsch & Borgwardt (ICLR 2019) that scores a trained neural network by running zero-dimensional persistent homology over each layer's complete weighted graph, transforming and sorting every weight to build the filtration — computed once, directly on trained weights, rather than tracked across a training trajectory. The third founding-era neural-network persistent-homology diagnostic examined in this vault's cluster, architecturally distinct from both Birdal's (random-sampling-over-time) and Gutiérrez-Fandiño's (no-sampling, full-network-every-step) constructions, yet sharing their absence of any structural-selection step.

References

• claim-rieck-2019-neural-persistence-computed-without-subsampling
• observation-hub-selection-risk-needs-incomplete-object-nn-diagnostics-read-complete-record
• entity-bastian-rieck · entity-persistent-homology-dimension